CoxKAN: Kolmogorov-Arnold Networks for Survival Analysis

Installation • Usage • Datasets • Reproducibility • Credits

CoxKAN leverages Kolmogorov-Arnold Networks for Interpretable, High Performance Survival Analysis.

• Installation: pip install coxkan
• Documentation: Read-the-Docs.
• Quick-start: tutorials/intro.ipynb

Repo Structure:

├── checkpoints/        # Results / checkpoints from paper
├── configs/            # Model configuration files
├── coxkan/             # CoxKAN package 
├── data/               # Data 
├── docs/               # Documentation
├── media/              # Figures used in paper
├── reprod/             # Reproducability instructions/code
├── tutorials/          # Tutorials for CoxKAN
|
# standard stuff:
├── .gitignore         
├── LICENSE          
├── README.md          
└── setup.py            

Installation

Pip

CoxKAN can be installed via:

pip install coxkan

Git

Alternatively, may desire the full codebase and environment that was used to produce all results in the associated paper:

git clone https://github.com/knottwill/CoxKAN.git 
cd CoxKAN
pip install -r reprod/requirements.txt

Please refer to reproducibility instructions in reprod/README.md.

Usage

Find tutorials in tutorials/ or Read-the-Docs

Example

from coxkan import CoxKAN
from coxkan.datasets import metabric 

df_train, df_test = metabric.load(split=True)

ckan = CoxKAN(width=[len(metabric.covariates), 1])

_ = ckan.train(
    df_train, 
    df_test, 
    duration_col='duration', 
    event_col='event',
    steps=100)

# evaluate model
ckan.cindex(df_test)
>>> 0.6441975461899737

CoxKAN Package

The coxkan/ package has 4 main components:

coxkan/
    ├── datasets/             # datasets subpackage
    ├── CoxKAN.py             # CoxKAN model
    ├── utils.py              # utility functions
    └── hyperparam_search.py  # hyperparameter searching

Datasets

Synthetic Datasets

coxkan.datasets.create_dataset makes it easy to generate synthetic survival data assuming a proportional-hazards, time-independant hazard function: $$h = h_0 e^{\theta(\mathbf{x})} \rightarrow T_s \sim \text{Exp}(h)$$ and uniform censoring distribution $T_c \sim \text{Uniform}(0, T_{max})$.

In the example below, we use a log-partial hazard of $\theta(\mathbf{x}) = 5 e^{-2(x_1^2 + x_2^2)}$ and a baseline hazard of $h_0 = 0.01$.

from coxkan.datasets import create_dataset

log_partial_hazard = lambda x1, x2: 5*np.exp(-2*(x1**2 + x2**2))
df = create_dataset(log_partial_hazard, baseline_hazard=0.01, n_samples=10000)

Clinical Datasets

5 clinical datasets are available with the coxkan.datasets subpackage (inspired by pycox). For example:

from coxkan.datasets import gbsg
df_train, df_test = gbsg.load(split=True)

You can decide where to store them using the COXKAN_DATA_DIR environment variable.

Dataset	Description	Source
GBSG	The Rotterdam & German Breast Cancer Study Group.	DeepSurv
METABRIC	The Molecular Taxonomy of Breast Cancer International Consortium.	DeepSurv
SUPPORT	Study to Understand Prognoses Preferences Outcomes and Risks of Treatment.	DeepSurv
NWTCO	National Wilm's Tumor Study.	Rdatasets
FLCHAIN	Assay of Serum Free Light Chain.	Rdatasets

Unfortunately, DeepSurv did not retain the column names. We manually restored the names by obtaining the datasets elsewhere and comparing the columns (then we can use the same train/test split):

• GBSG: https://www.kaggle.com/datasets/utkarshx27/breast-cancer-dataset-used-royston-and-altman
• SUPPORT: https://hbiostat.org/data/repo/support2csv.zip
• METABRIC: https://www.kaggle.com/datasets/raghadalharbi/breast-cancer-gene-expression-profiles-metabric

Genomics Datasets

We curated 4 genomics datasets from The Cancer Genome Atlas Program (TCGA). The raw or pre-processed data is available by request - please email me at knottenbeltwill@gmail.com.

Two of the datasets (GBMLGG, KIRC) were the unaltered datasets used in Pathomic Fusion

Dataset	Description	Source
STAD	Stomach Adenocarcinoma.	TCGA
BRCA	Breast Invasive Carcinoma.	TCGA
GBM/LGG	Merged dataset from two types of brain cancer: Glioblastoma Multiforme and Lower Grade Glioma.	Chen et al.
KIRC	Kidney Renal Clear Cell Carcinoma.	Chen et al.

Reproducibility

All results in the associated paper can be reproduced using the codes in reprod/. Please refer to the instructions in reprod/README.md.

Credits

Special thanks to:

• All authors of Kolmogorov-Arnold Networks and the incredible pykan package.
• Håvard Kvamme for pycox and torchtuples.